OpenAI says its AI solved 10 long-standing math problems

In short: AI systems are starting to solve math problems that researchers have worked on for decades, and that is changing how mathematicians think about their field.

What's going on

OpenAI says an internal, unreleased AI model it calls Astra produced solutions to 10 long-standing mathematics problems. These include topics that matter for real-world tech, like packing “spheres” in higher dimensions (which links to how data can be stored and sent) and improving error-correcting codes (methods that help recover information from noisy signals, like a scratchy radio transmission).

OpenAI published more than 250 pages of papers explaining the results, plus additional write-ups on how the ideas were developed. The company also said the proofs were checked with Lean, a tool that verifies math proofs step by step (like a spellchecker, but for logic).

The Verge reports that many mathematicians are excited, but also uneasy. One dispute focused on a result about “non-sofic groups,” a hard, abstract question about infinite mathematical structures. Researchers said OpenAI’s original wording made it sound like there had been little human progress, and it later changed the language without a clear correction note. OpenAI told The Verge it updated the post to better acknowledge prior work.

What to watch

Cost and access could become a dividing line. OpenAI estimated about $2,000 in usage fees to generate the 10 solutions at current prices, which can be high for a field that often runs on small budgets. Researchers also worry that closed, company-controlled systems could shape what gets studied, who gets credit, and whether students can still learn by struggling through problems that AI might now solve quickly.

Source: The Verge AI